AI Agents (MCP)

MCP uses your existing platform account. It does not create accounts on its own.

Your AI agent or MCP client likely supports our OAuth mechanism, which lets you authorize it through a browser session. The login tool lets the client request authentication with your platform email and password when OAuth is not used or when the client needs to refresh credentials.

You can authorize MCP HTTP requests with an API key from User settings. Clients that cannot use OAuth should send it as Authorization: Bearer. A one-day access token from the login endpoint also works until it expires.

What the server offers

• Instructions. get_mcp_instructions returns the standing rules: the design playbook and the per-tool hints.
• Docs. get_docs is the docs reader. Use mode=manifest to get the whole tree in one call, mode=content for one page, and mode=bundle to fetch several pages plus their dependencies at once.
• Tools. Every registered tool has a page under MCP tools, with inputs, the REST endpoints it calls, and the admin screen that does the same work.
• Design catalogs. Block schemas come from the live catalog, so get_docs is the only source you should insert block types from.

Reading efficiently

One get_docs call with mode=bundle replaces the usual discovery loop:

• paths names the pages you need, for example design/page and objects/lifecycles.
• include adds dependencies: body, ancestors, children, shared-concepts, blocks, lifecycles, related-tools, related-endpoints.
• depth controls how many levels of children come back.
• budgetBytes caps the response. Anything that does not fit is listed in omitted instead of being cut silently.

Connect

Connect and authenticate covers OAuth, API keys, and the login tool. The REST equivalent of every tool is in the API endpoints catalog.